Though there are some places where the wording is cautious, there is a strong desire in this journal article, and the associated news articles, to make the interpretation that having a better attitude toward aging will result in better objective aging outcomes. That might be true, and seems likely to at least be true at a low effect size. But the “association” or correlation between aging satisfaction (the measuring of attitude toward aging) and objective aging outcomes could easily be an instance of people having a sense of their own health status that goes beyond easily tallied objective outcomes and having a worse attitude toward aging because they already know their health is poor.

I don’t see how the effect of attitude toward aging on objective aging outcomes can be identified without a randomized controlled trial of an intervention trying to improve attitudes toward aging. Note that getting a placebo here is a little tricky. It requires something like an intervention to change an attitude thought to be irrelevant to aging outcomes.

If one considers underlying health status to be distinct from easily tallied health outcomes, then one could see this as “Cousin Causality” from underlying [health status] affecting both [aging satisfaction] and [later easily tallied bad health outcomes]. A virtue of seeing this as an instance of cousin causality (sometimes called “third-factor causality”) as opposed to reverse causality is that it makes clearer that the aging satisfaction being measured earlier in time than the easily tallied health outcomes doesn’t mean the aging satisfaction caused the health outcomes. The ancestor factor affecting both can easily be earlier in time than both. Then on one branch we are saying that poor underlying health begets bad easily tallied health outcomes, which seems like a truism.

In the conclusion, the authors Julia S. Nakamura, Joanna H. Hong, and Jacqui Smith write:

… there is potential for confounding by third variables. However, we addressed this concern by implementing a longitudinal study design, robust covariate adjustment, and E-value analyses.

“Longitudinal study design” could mean two things: one, having aging satisfaction measured before easily tallied health outcomes. That doesn’t solve a problem of “confounding by third variables,” which is another way of describing cousin causality, for the reason I said above: the third ancestor variable can easily be before both aging satisfaction and easily tallied health outcomes. The other thing “longitudinal study design alludes to is a focus on changes in aging satisfaction. But changes in aging satisfaction can easily be due to changes in underlying health status, and everything I say above goes through. “Robust covariate adjustment” doesn’t solve the problem of confounding because subjective health measures contain information that goes beyond all of the other health measures in the Health and Retirement Study (HRS); research shows that subjective health measures add predictive power for later easily tallied health outcomes beyond other covariates in the HRS. I am suggesting that aging satisfaction has to be considered as being analogous to the subjective health measures in the HRS. Note that doesn’t mean controlling for subjective health would solve the problem either, because subjective health is surely measured with error. (See “Adding a Variable Measured with Error to a Regression Only Partially Controls for that Variable.” Having aging satisfaction as a second measure of subjective health ought to increase predictive power over only one measure of subjective health. Finally, the “E-value analyses” simply say that the story for why there might be confounding would have to be one that makes a lot of confounding plausible. I think that is satisfied by the story I give above.